AI is getting cheaper, but needs more resources
Artificial intelligence is becoming more accessible for businesses and private users. Cheaper models and open weights make computation less costly per unit of work, but overall demand for compute power is rising at the same time.
This trend was outlined briefly by Grayscale’s head of research Zak Pandle in one of his talks and he warned on social media about the risk that growth in AI usage could outpace improvements in model efficiency.
Why this became possible
The main reason for the shift in the cost–demand balance is the emergence of autonomous AI agents. Unlike conventional chatbots that respond to individual queries, agents can carry out sequences of actions over extended periods. They analyse data, process documents and prepare outputs without constant human intervention.
At the same time, competition among model developers and access to open weights have driven down the cost of computation per request. For example, according to Anthropic, its Sonnet model costs up to 30% less to perform a single task. Lower per‑unit prices encourage wider adoption of AI‑based services and thereby increase the total number of operations performed.
Why demand for compute is surging right now
The combination of cheaper tokens and the emergence of autonomous agents has created a multiplying effect. An agent can run through large volumes of text and data continuously, driving token consumption many times higher than a typical human‑to‑chatbot interaction.
Analytics confirm rapid growth in load. According to OpenAI, OpenRouter and Similarweb in a piece by the crypto arm of Andreessen Horowitz a16z, autonomous agents use almost five times more tokens than humans. In addition, their overall consumption has grown roughly 14× since February 2026. Grayscale cites estimates that, depending on the task, agent processes can require 5−50 times more tokens than a standard chatbot interaction.
Who will be affected by rising demand for compute
First and foremost, this concerns infrastructure providers. Digital services rely on the electricity powering data centres and on graphics processing units (GPUs). If demand for compute grows faster than model efficiency and infrastructure expansion, competition for resources will intensify.
According to a forecast from Goldman Sachs cited by Pandle, token consumption could increase 24‑fold by 2030. That means that even with further reductions in per‑operation costs, total compute needs will rise. As an illustration, Pandle pointed to CoreWeave, which has signed new compute contracts at higher prices.
Who will benefit
In a resource‑constrained environment, the winners will be companies that control access to the necessary infrastructure. These are data‑centre providers, GPU suppliers and large cloud operators that can scale capacity faster than others.
The market will also stratify, with the most active customers consuming far more tokens. Analysts note that the busiest firms generate more than eight times the tokens of a typical enterprise. The spread of agents also raises memory requirements and changes approaches to building traditional automation systems.
What this means for businesses and users
For businesses, it is a signal to reassess infrastructure investments. Companies will need to forecast not only the cost of a single transaction but the aggregate load that autonomous agents will create. At the same time, cheaper AI will expand its use in analytics, automation and decision‑support tools.
For end users, this could mean broader access to AI‑based services but also possible increases in the price of specialised services if resources become scarce. In practice, market changes will be determined by who can offer affordable compute and flexible terms of cooperation.
